Problem
Tinkoff operators serve clients in an in-house CRM and must pick one of tens of thousands of service procedures, for example "debt amount on a credit card". The tool must fit into the operator interface. The customer first asked for a model that reads the client's phrases, which the team judged too large for the one-month deadline. The customer's constraints: at most 6 recommended procedures, a real-time service, and access rights that hide some procedures from some operators.
Example from the article: a client asks how much they owe. The operator launches "debt amount", and the next useful procedures are partial early repayment or cash withdrawal.
Solution in brief
Split the task in two. A cold-start list for the first procedure comes from counting the most frequent first procedures, per operator or per operator tag group. Later procedures come from a probabilistic transition tree over the history of launched procedures. Precomputed lists live in Redis, and a Lua script filters them, so no model server is needed.
Questions this case answers
- How to recommend the next action from a very large catalog in real time under access restrictions and a one-month deadline
- How to serve precomputed recommendations at low latency with minimal infrastructure
- How to recommend the next service procedure to a support operator in real time when there are tens of thousands of procedures?
- How to build a cold-start list of procedures before the operator has launched anything in a call?
Tools and methods
- Top Pop by tags and Personal Top Pop — cold-start lists from counts of first launched procedures
- Transition tree over procedure chains — next-procedure lists with fallback to shorter histories
- Redis with Lua scripts — storage and filtering of precomputed lists as an inference server
- FastAPI — thin service layer with separate endpoints for the two models
Google or a one-time email code. No password.
